Every frequency counter we own agrees that the oscillator in front of me is running 0.014% flat of perfect A440. By the strictest reading of the spec sheet, this is a defect. By every other reading, including the one that matters to your ears, this is a voice.
This note is about the distance between those two readings. It is also about the month we spent proving, to ourselves more than to anyone else, that the 0.02% tolerance we print on the box is not a hedge. It is a design decision with a sound attached to it, and you should be able to hear the argument before you buy it.
What drift actually is
An analog oscillator keeps its pitch because a capacitor charges through a current source at a predictable rate. "Predictable" is doing heavy lifting in that sentence. The current source depends on a reference voltage, which depends on temperature, which depends on how long the instrument has been on, whether the sun is on the case, and whether the person next to you is breathing on it. A digital oscillator does not have this problem, because a digital oscillator does not have a capacitor in the signal path. It has a clock and mathematics, and neither cares about the sun.
Drift, then, is the slow wander of an analog pitch away from where you put it. In a cheap design, the wander is wide, unpredictable, and audible as a synth that is out of tune with itself. In a careful design, the wander is narrow, slow, and correlated, which is a completely different animal. Ours wanders at most 0.02% over a full studio day and wanders together, voice with voice, the way an ensemble moves together.
The measurement we did not want to publish
In March we set up the comparison that settled this for us. Four studios in Berlin lent us bench time, and we shipped each one a HELIOS-8 prototype, a calibration certificate, and a notebook. The task was simple: record the same pad patch at every hour of a working day, twice, and log the counter reading at each pass.
The result is the chart below. Each line is one voice of one unit across one studio day. The spread at any moment is the "disagreement" between voices; the slow rise on each line is the morning warm-up. Both are real. Both are inside the printed tolerance.
Read the chart two ways and you will see both readings at once. The metrologist reads it and says: error, within bounds. The musician reads it and says: eight lines that started as one chord and became eight slightly different opinions about what that chord is. The second reading is the one we print boxes for.
Why "locked" is not the ceiling
It is tempting to say the answer is obvious: lock the tuning, keep the precision, everyone goes home. We tried it. The firmware has a "locked" mode that trims the voices against the calibration table every 30 seconds, and in that mode the chart above becomes eight flat lines. It measures better and sounds worse, and we can now say why with numbers instead of vibes.
When voices wander, the upper harmonics of the chord move at slightly different rates, which means the beating pattern inside the chord evolves. Humans read that evolution as space, dimension, and, honestly, as other people. When the voices are locked, the beating pattern is frozen, and the pad sounds like a photograph of a pad. In blind listening across the four studios, seven of nine engineers picked the drifting take as "wider" and "more expensive sounding," and one picked the locked take as "more precise," which is exactly the word for it and exactly not the word we want on a record.
The compromise, written down
The firmware ships with three drift settings and one honest lie detector. "Locked" is the flat-line mode for when you are tuning to an orchestra. "Gentle" is the default, holding voices within 0.008% of each other, which is inside the ensemble band of the chart. "Wide" lets each voice use the full printed 0.02%, which is a sound, not a mistake. The honest lie detector is a spec: if a module drifts outside its printed tolerance, that is a warranty repair, not a vibe. The certificate in your box is the boundary between the two, and we defend it.
Method, so you can argue with us
- Frequency counter: calibrated universal counter, 10-digit resolution, oven-stabilized reference, warm 2 hours before session start.
- Units: four HELIOS-8 revision B prototypes, serials 0008 through 0011, calibration certificates dated 2026-03-02.
- Procedure: all voices patched to a common drone, counter logged on the hour, room temperature logged alongside, first pass at power on, last pass at 8 hours.
- Studios: four rooms in Berlin with wildly different heating systems, which turned out to be the most useful variable in the whole study.
- Blind listening: 9 engineers, 14 pairs of takes, locked versus gentle versus wide, ABX method, no time limit.
Corrections. An earlier version of this note said the blind listening used "eleven engineers." It was nine; the extra two were our own bench and should not have counted. The chart is unchanged.
Filed under: oscillators, tuning, tolerance. Comments and counter-measurements welcome at the bench address on the contact page. The best counter-measurement we have ever received became its own lab note.